Hall of Fame Probability Model for NBA Coaches
In the following, the model used for estimating NBA Coaches' Hall of Fame probabilities is presented in full detail.
Data
The data set consists of all head coaches listed in the Basketball-Reference coach directory, covering the BAA (1946–49), the NBA (1949–2026) and the ABA (1967–76). Career totals of regular-season games and wins were aggregated from season-level coaching tables, and championships and Finals appearances were attributed, for each finalist team-season, to the coach who led the team in that season's playoffs. Coach of the Year awards from both leagues were included. NCAA records were obtained from Sports-Reference and linked to professional careers by name and date of birth. Out of the professional coaches, 118 were linked to an NCAA record, yielding NCAA wins, tournament appearances, Final Fours and national titles.
The response \(y_i = 1\) indicates induction into the Naismith Memorial Basketball Hall of Fame in the Coach category. Persons inducted as players or contributors who also coached are coded \(y_i = 0\). The estimation sample is restricted to coaches eligible under the Hall’s rules, i.e., fully retired for at least five years or active with at least 25 seasons as a head coach. This yields \(n = 334\) coaches, of whom 31 (9.3%) have been inducted.
Initial model
The probability of induction \(p_i = \Pr(y_i = 1 \mid \mathbf{x}_i)\) is modelled by logistic regression,
\[ \operatorname{logit}(p_i) \;=\; \log\frac{p_i}{1-p_i} \;=\; \beta_0 + \sum_{j=1}^{13} \beta_j x_{ij}, \] \[ p_i \;=\; \frac{1}{1+\exp\!\bigl(-\beta_0-\sum_{j}\beta_j x_{ij}\bigr)} , \]where the thirteen potential independent variables presented in Table 1, describe the length and success of the professional career, individual recognition, and the college career. For coaches without an NCAA record the NCAA winning percentage is undefined. It is set to zero and accompanied by the NCAA indicator, so that its coefficient pertains only to coaches with college experience.
| Variable | Description |
|---|---|
| NBA Start year | First season as an NBA/ABA head coach |
| NBA Seasons | NBA/ABA seasons with at least one game coached |
| NBA Games \( G_i \) * | NBA/ABA regular-season games coached |
| NBA Win % | NBA/ABA regular-season winning percentage |
| NBA Championships \( C_i \) * | NBA/ABA (and BAA) championships |
| NBA Finals | NBA/ABA Finals appearances |
| COY | Coach of the Year awards (NBA and ABA) |
| NCAA indicator | Indicator of NCAA head-coaching experience |
| NCAA games | NCAA games coached |
| NCAA wins \( W_i \) * | NCAA games won |
| NCAA win % | NCAA winning percentage (0 if no record) |
| NCAA tournaments | NCAA tournament appearances |
| NCAA Final Fours | NCAA Final Four appearances |
| NCAA titles \( T_i \) * | NCAA national championships |
Variable selection
The full model exhibited severe multicollinearity. Seasons and games coached are almost perfectly correlated (\(r = 1.00\), which produced opposite-signed coefficients that merely contrast season lengths across eras. Similarly, championships and Finals appearances are strongly correlated (\(r = 0.91\)), as are the NCAA indicator and NCAA winning percentage (\(r = 0.95\)) and the NCAA volume and success measures among themselves (\(0.71 \le r \le 0.86\)).
Bidirectional stepwise selection under the Bayesian information criterion retained games, championships, NCAA games and NCAA titles. Candidate models with small variations were then compared by repeated ten-fold cross-validation using the area under the ROC curve (AUC), the Brier score and the logarithmic loss. Replacing NCAA games by NCAA wins improved the out-of-sample log-loss and yielded a more interpretable variable, which led to the final specification
\[ \operatorname{logit}(p_i) \;=\; \beta_0 + \beta_1 G_i + \beta_2 C_i + \beta_3 W_i + \beta_4 T_i . \]Because NCAA titles nearly separate inductees from non-inductees, maximum-likelihood estimates are biased away from zero and the fitted probabilities are overconfident. The final model is therefore estimated by Firth’s penalized likelihood,
\[ \ell^{*}(\boldsymbol\beta) \;=\; \sum_{i=1}^{n}\Bigl[\,y_i \log p_i + (1-y_i)\log(1-p_i)\Bigr] \;+\; \tfrac12 \log\bigl|\mathbf{I}(\boldsymbol\beta)\bigr|. \]Comparison of the full and final models
Table 2 compares the full model with the final four-variable model. The likelihood-ratio test of the nine excluded variables is not significant, and both information criteria favour the reduced model, the Bayesian criterion decisively so. Based on the out-of-sample comparison, the full model overfits, and the final model improves all three cross-validated measures, with Firth’s correction lowering the log-loss further by about 17%. The Hosmer–Lemeshow test on deciles of fitted probability gives no evidence of miscalibration in the final model and the expected number of inductees, \(\sum_i \hat p_i = 31.4\), matches the observed 31.
| Full model | Final model (ML) | Final model (Firth) | |
|---|---|---|---|
| Parameters | 14 | 5 | 5 |
| Residual deviance | 47.39 | 60.12 | – |
| AIC | 75.39 | 70.12 | – |
| BIC | 128.75 | 89.17 | – |
| Cross-validated AUC | 0.939 | 0.954 | 0.955 |
| Cross-validated Brier score | 0.038 | 0.030 | 0.030 |
| Cross-validated log-loss | 0.210 | 0.146 | 0.122 |
Final model
The Firth estimates are reported in Table 3, and the fitted model reads
\[ \hat p \;=\; \Bigl[\,1+\exp\!\bigl(6.129 - 0.00293\,G - 1.916\,C - 0.00711\,W - 2.847\,T\bigr)\Bigr]^{-1}. \]All four effects are positive and significant at the 1% level. Because the model is non-linear in probability, the effect of each variable depends on the rest of the résumé, and it is most naturally described through predicted probabilities (Table 4).
Variable \(\hat\beta\) Std. error \(z\) \(p\)-value Intercept −6.129 0.856 −7.16 < 0.001 NBA/ABA games \(G\) 0.00293 0.00061 4.83 < 0.001 NBA/ABA championships \(C\) 1.916 0.440 4.35 < 0.001 NCAA wins \(W\) 0.00711 0.00196 3.63 < 0.001 NCAA titles \(T\) 2.847 0.906 3.14 0.002
Games coached No championship One championship Two championships 0 0.2 (0.0–1.2) 1.5 (0.3–6.2) 9.1 (1.7–37.3) 500 0.9 (0.3–3.2) 6.0 (2.2–15.4) 30.3 (9.2–65.0) 1,000 3.9 (1.5–10.2) 21.7 (10.9–38.4) 65.3 (33.2–87.6) 1,500 15.0 (5.5–34.7) 54.5 (31.4–75.8) 89.0 (65.4–97.2) 2,000 43.2 (15.2–76.4) 83.8 (56.6–95.4) 97.2 (85.5–99.5)
A newly appointed head coach with no games, titles or college record has a baseline induction probability of 0.22%. Coaching an additional 100 NBA games raises this only to 0.29%, but the increments compound: with no championship, the probability is 3.9% after 1,000 games and 43% after 2,000 games, and averaged over the eligible coaches an additional 100 games raises the induction probability by 0.85 percentage points.
A championship has a substantially larger effect: it raises the baseline probability from 0.22% to 1.5%, and for a coach with 1,000 games from 3.9% to 21.7%, an average increase of 7.7 percentage points across the sample. Equivalently, the induction probability reaches one half after approximately 2,090 games without a championship, 1,440 games with one and 790 games with two.
Among college credentials, 100 NCAA wins increase the probability by 2.2 percentage points and an NCAA title by 14.3 percentage points on average; for example, a coach with 600 NCAA wins and no NBA career has a probability of 13% without a national title and 73% with one. The NCAA title effect is estimated from a small number of coaches and carries a correspondingly wide confidence interval. Finally, the model captures only the quantifiable components of a coaching career and contributions such as the integration of the game, which explains the induction of, e.g., John McLendon, lie beyond its scope.
